"Ask your site": the history that answers, and the questions it refuses

An HSE manager usually knows what happens on their site. What they lack is the evidence, the duration and the recurrence: how many times, for how long, at what hours, and since when. Camly records every situation as a dated episode, with the image, the camera and the zone. The history becomes queryable in plain language, directly, without exports or spreadsheets. And it refuses, by design, to answer about a person.

What you can ask

The chat relies only on what Camly has recorded: episodes, alerts, the checks that were carried out. It does not look at live raw video to improvise an answer. A few questions a site asks in an ordinary week:

• “How many times was emergency exit B blocked this month, and for how long?”
• “What is at the charging station right now?”
• “Which zone had the most episodes this week?”
• “Show me the episodes at dock 3 between 6 and 8 this morning.”
• “Was fire aisle 2 cleared after yesterday’s alert?”

The answer is written, in French on French sites, and it comes with what it relies on: the list of episodes concerned, their start and end times, their duration, and the evidence images.¹ Each episode can be opened from the answer. If the history holds nothing on the subject, the chat says so rather than inventing.

Question	What the answer contains
“How many times… this month?”	The number of episodes, cumulative duration, average duration, the most frequent hours
“What is… right now?”	The latest finding on the zone, its time, the image
“Which zone… this week?”	A ranking by zone, with the number of episodes and the duration
“Show me… between 6 and 8”	The episodes in the range, in order, with the images
“Was it cleared after…?”	The time the episode closed, the delay since the alert

What it refuses, and why that is a feature

Some questions get a refusal, not an approximate answer:

• any question about a named or identifiable employee: “what was Karim doing at 10?”;
• any question about individual productivity: “who is slowest at dock 2?”;
• any attribution: “who left this pallet in front of the exit?”.

Camly does no facial recognition, no biometrics, no identification. Findings are about places and situations; people are blurred on the images kept.¹ The chat applies the same rule: it speaks of zones, states and durations, never of individuals. This is not a technical limit that a better-phrased question would get around. It is a product rule.

For the HSE manager, this is what makes the tool presentable to the works council. The information pack describes exactly what is recorded, what is not, and what the chat refuses. A tool that could answer “who left this pallet” would end up doing so, one day, under pressure, and that is precisely the kind of use the CNIL sanctions and the EU AI Act regulates.² A tool that cannot protects the HSE manager as much as the staff.

Three concrete uses

Preparing the monthly report. Before, the month was reconstructed from memory and a few phone photos. Now the question “which situations came back most often this month, by zone?” gives the list, the durations and the recurrences. Camly’s near-miss report is built on those episodes; the chat is there to check and complete it.

Answering an inspector or an insurer. The standard question of a visit is “since when has this exit been clear, and how many times was it blocked?”. The answer exists, dated, with the images. The site no longer has to say “we are careful”; it shows.

Checking a corrective action. After moving a buffer storage area, you want to know whether the fire aisle stays clear. “How many episodes on aisle 3 since the 15th?” settles it in one sentence, and the trend is readable in the history.

What the chat is not

It does not replace the day replay or the dashboard: it is faster on a precise question and weaker on an overview. It only sees what the site’s tasks watch: if no task covers dock 4, there is nothing to say about dock 4. And it does not judge: it describes a state, a duration, a recurrence. The decision stays with the site.

If you want to see what your site would answer, a thirty-day pilot on the cameras of one site is enough for the history to have something to say.

Sources

1. Camly AI, brochure “Camly for warehouses”, pilot edition, October 2026: alerts and episodes, chat, privacy and sovereignty.
2. CNIL, deliberation of 27 December 2023, Amazon France Logistique (€32M); Regulation (EU) 2024/1689 on artificial intelligence, Annex III, point 4. cnil.fr